Full Catalog

Benchmarks

Every AI biology and biomedical benchmark we track, with saturation status, scores, and successor information.

222 of 222 benchmarks
NameDomainYearTypeStatusTime to SaturationBest ScoreHuman BaselineSuccessor
MoleculeNetMolecular/Drug Discovery2018Classification/RegressionNearing96 moAUROC 0.85-0.95—TDC, WelQrate
TDC ADMET (22 leaderboards)Molecular/Drug Discovery2021Classification/RegressionActive—Varies by task—TDC-2
TDC DrugComboMolecular/Drug Discovery2021RegressionActive—Varies——
TDC DockingMolecular/Drug Discovery2021GenerationActive—Varies——
TDC DTI DG GroupMolecular/Drug Discovery2021RegressionActive—Varies——
TDC Clinical Trial OutcomeMolecular/Drug Discovery2023ClassificationActive—Varies——
TDC-2 (Single-cell DTI)Molecular/Drug Discovery2024ClassificationActive—Varies——
TDC-2 Protein-Peptide BindingMolecular/Drug Discovery2024RegressionActive—Varies——
GuacaMolMolecular/Drug Discovery2019GenerationSaturated24 moNear-perfect on simple goals—PMO, TARTARUS, DrugGym
MOSESMolecular/Drug Discovery2019GenerationSaturated48 moHigh validity/uniqueness—MolScore, TARTARUS
PMOMolecular/Drug Discovery2022OptimizationActive—Varies——
TARTARUSMolecular/Drug Discovery2023Generation/RLActive—Low success on hard objectives——
DrugGymMolecular/Drug Discovery2024Agent WorkflowActive—N/A (simulator)——
DOCKSTRINGMolecular/Drug Discovery2022Regression/GenerationActive—Varies——
MolScoreMolecular/Drug Discovery2024Meta-frameworkUnknown—N/A——
MolGenBenchMolecular/Drug Discovery2025Structure-based GenerationActive—Varies——
ToxiMolMolecular/Drug Discovery2025GenerationActive—Low success rates——
Tox21Molecular/Drug Discovery2014ClassificationUnknown—AUC ~0.85——
PolarisMolecular/Drug Discovery2024Classification/RegressionActive—Varies (blind test)——
WelQrateMolecular/Drug Discovery2024ClassificationActive—Varies——
PharmaBenchMolecular/Drug Discovery2024Classification/RegressionActive—Varies——
TOXRICMolecular/Drug Discovery2023ClassificationActive—Varies——
ZINCMolecular/Drug Discovery2015Regression/GenerationActive—Varies—ZINC20, ZINC-22
PDBbindMolecular/Drug Discovery2004RegressionNearing264 moPearson R >0.86 (standard); <0.60 (CleanSplit)—PDBbind CleanSplit, PLINDER
PDBbind CleanSplitMolecular/Drug Discovery2025RegressionActive—Varies (leak-free)——
PLINDERMolecular/Drug Discovery2024Docking/RegressionActive—DiffDock drops 38%->15% with leak control——
CASF-2016Molecular/Drug Discovery2016Scoring/Ranking/DockingNearing120 moPearson R ~0.86—PDBbind CleanSplit
DAVISMolecular/Drug Discovery2011RegressionNearing180 moCI ~0.903——
KIBAMolecular/Drug Discovery2014RegressionActive—CI ~0.898——
CPI2MMolecular/Drug Discovery2024RegressionActive—Baseline——
PoseBustersMolecular/Drug Discovery2023Pose ValidationActive—~75% PB-Valid——
PoseBenchMolecular/Drug Discovery2024Pose PredictionActive—~68% (PB); ~33% (DockGen-E)——
DockGenMolecular/Drug Discovery2024Pose PredictionActive—~24-33% top-1 RMSD<2A——
PoseXMolecular/Drug Discovery2025Pose PredictionActive—AI > physics-based (post-relaxation)——
LIT-PCBAMolecular/Drug Discovery2020Virtual ScreeningActive—Varies—MF-PCBA
DUD-EMolecular/Drug Discovery2012Virtual ScreeningSaturated96 moBiased high AUC—LIT-PCBA, MF-PCBA, WelQrate
MF-PCBAMolecular/Drug Discovery2024Virtual ScreeningActive—Baselines established——
CrossDocked2020Molecular/Drug Discovery2020Scoring/PoseActive—R 0.612——
USPTO-50KMolecular/Drug Discovery2016RetrosynthesisNearing120 mo65% top-1~48.2% avg (forward)—
USPTO-MITMolecular/Drug Discovery2017Reaction PredictionNearing108 mo>90% top-1 (forward)——
Schneider 50KMolecular/Drug Discovery2015Reaction ClassificationSaturated72 mo>99% (RXNFP)—USPTO 1K TPL
TOMG-BenchMolecular/Drug Discovery2024LLM GenerationActive—Fine-tuned Llama > GPT-3.5 by 46.5%——
MolQuestMolecular/Drug Discovery2026Agent WorkflowActive—SOTA ~50%; most <30%——
ChemBenchMolecular/Drug Discovery2024MCQ/Free-formNearing24 moClaude 3.5 Sonnet > best human chemistsExpert chemists—
QM9Molecular/Drug Discovery2014RegressionNearing144 moNear chemical accuracy—PCQM4Mv2
QM7/QM7bMolecular/Drug Discovery2012RegressionSaturated96 moChemical accuracy achieved—QM9
QM8Molecular/Drug Discovery2015RegressionNearing132 moNear saturation—QM9
PCQM4Mv2Molecular/Drug Discovery2021RegressionActive—MAE 0.0719 eV——
CASP (Single-Chain)Protein1994Structural PredictionSaturated312 moGDT-TS 92.4~90 (experimental)CASP complexes/RNA tracks
CASP16 (Complexes)Protein2024Structural PredictionActive—Varies; AB-Ag major failure area——
CASP16 (RNA)Genomics2024Structural PredictionActive—No TM-score >0.8 for novel RNAs——
CASP16 (Protein-Ligand)Molecular/Drug Discovery2024Binding AffinityActive—Max Kendall tau 0.42——
CAMEOProtein2013Structure EvaluationSaturated144 moAlphaFold dominant—CAMEO complexes/ligands/peptides
CATHProtein1997ClassificationNearing348 moHigh accuracy (known folds)——
TAPEProtein2019Multi-taskSaturated48 moNear-ceiling on most tasks—ProteinGym, PEER, FLIP
PEERProtein2022Multi-taskActive—Varies across 17 tasks——
ProteinGymProtein2023RegressionActive—Spearman ~0.52——
ProteinBenchProtein2024Multi-taskActive—Varies; drops >20% at 500+ residues——
FLIPProtein2021RegressionNearing60 moVaries—FLIP2
FLIP2Protein2026RegressionActive—Varies——
FoldBenchProtein2025Structural PredictionActive—Varies by task——
CAPRIProtein2001Docking/InteractionActive—Varies——
ProteinInvBenchProtein2023GenerationActive—Recovery ~66%——
CAFAProtein2010Function PredictionActive—Fmax 0.4-0.6——
DB5.5 (Docking Benchmark)Protein2003DockingActive—Top-10 success ~38%——
SKEMPI 2.0Protein2019RegressionActive—Pearson R ~0.7-0.8——
PPB-AffinityProtein2024RegressionActive—Varies——
SAbDab (CDR design)Protein2014GenerationActive—CDR-H3 AAR ~40-50%; RMSD ~2.5-3.5A——
AbBiBenchProtein2025Antibody DesignActive—Varies——
Mega-Scale Stability DatasetProtein2023RegressionActive—PCC 0.72 (ThermoMPNN)——
S669 (Stability blind test)Protein2020RegressionActive—PCC ~0.43-0.67——
EC-BenchProtein2025ClassificationActive—Varies——
MotifBenchProtein2025GenerationActive—RFdiffusion solves ~16/30——
SHAPESProtein2025GenerationActive—Most models fail on loops/mixed——
PFMBenchProtein2025Multi-taskActive—Varies across 38 tasks——
LiveProteinBenchProtein2025Multi-taskActive—Varies——
ATOM3DProtein2021Multi-task (3D)Active—Varies across 8 tasks——
OGB Protein TasksProtein2020Classification/Link PredictionActive—Varies——
LRGB Peptides-funcProtein2022ClassificationActive—Varies (AP)——
LRGB Peptides-structProtein2022RegressionActive—Varies (MAE)——
RNA-PuzzlesProtein2012Structural PredictionActive—Best RMSD 3-7AHuman experts outperform servers—
BEACON (RNA Tasks)Genomics2024Multi-task (13 RNA tasks)Active—PLMs surpass SOTA on 8/13——
rnaglibProtein2025Multi-task (RNA 3D)Active—Varies——
ATLAS MD Ensemble DatasetProtein2024Conformational DynamicsActive—Varies——
PeptoneBenchProtein2025Conformational DynamicsActive—BioEmu and PepTron lead——
PDFBenchProtein2025Protein Design from FunctionActive—Varies across 16 metrics——
SafeProteinBiosecurity2025Red-teamingActive—Varies——
GenomicBenchmarksGenomics2022ClassificationSaturated18 mo95%+ accuracy—GUE, BEND
GUEGenomics2023ClassificationActive—F1 0.3-0.95—DART-Eval, GenBench
BENDGenomics2023Classification/RegressionActive—Varies——
NT-18 BenchmarkGenomics2023ClassificationActive—MCC 0.974 (promoter); 0.983 (splice)—NTv3
Gene-MTEBGenomics2025Classification/ClusteringActive—0.59 average——
NullsettesGenomics2025Zero-shot PredictionActive—Most models fail——
Evo / Evo 2Genomics2024Multi-taskActive—BRCA1 >90% AUROC (Evo 2)——
METAGENE-1Genomics2025Classification/ClusteringActive—MCC 92.96 (pathogen)——
DART-EvalGenomics2024Multi-taskActive—DNALMs inconsistent——
OmniGenBenchGenomics2025Multi-taskActive—Varies across 123+ datasets——
DNALongBenchGenomics2025Classification/RegressionActive—Expert models lead——
GenBenchGenomics2024Multi-taskActive—Varies across 43 datasets——
NABenchGenomics2025Fitness PredictionActive—Varies——
Genomics Long-Range BenchmarkGenomics2024Classification/RegressionActive—Large gaps in VEP——
DeepSEAGenomics2015ClassificationNearing132 moAUC 0.958—Sei, AlphaGenome
SeiGenomics2022ClassificationActive—AUROC 0.972; AUPRC 0.409—AlphaGenome
AlphaGenomeGenomics2025Multi-taskActive—25/26 VEP evaluations SOTA——
EnformerGenomics2021Gene Expression PredictionActive—Cross-gene CAGE Pearson 0.85~0.94 (experimental replicate ceiling)Borzoi, AlphaGenome
BorzoiGenomics2024Gene Expression PredictionActive—524kb input, 32bp resolution—AlphaGenome
SpliceAIGenomics2019ClassificationNearing84 moAUPRC 0.98; ~95% top-k—OpenSpliceAI, AlphaGenome
CAGIGenomics2011Variant InterpretationActive—Varies——
TraitGymGenomics2025Variant PredictionActive—CADD/GPN-MSA best for disease traits——
ClinVar (coding variants)Genomics2013ClassificationNearing156 moAUC ~0.95——
CAMI (Metagenomic)Genomics2017Classification/AssemblyActive—Good at genus, poor at strain—CAMI III
DNABERT evaluation tasksGenomics2021ClassificationNearing60 moF1 0.940 (promoter); MCC 0.871 (splice)—DNABERT-2, GUE
CRISPR on-target (CRISPRon)Genomics2021RegressionNearing60 moSpearman ~0.80Ceiling ~0.85-0.90 (biological noise)—
CRISPR off-target (CIRCLE-seq)Genomics2017ClassificationActive—AUROC 0.977 (CCLMoff)——
PRIDICT 2.0 (Prime Editing)Genomics2024RegressionActive—Spearman R=0.85——
DeepPrimeGenomics2023RegressionActive—Approaching strong performance for specific edits——
scIBVirtual Cell2020IntegrationNearing72 moscANVI ~0.8—CZI cz-benchmarks
CZI cz-benchmarksVirtual Cell2024Classification/PredictionActive—Varies——
Open Problems (single-cell)Virtual Cell2021Multi-taskActive—Varies across 12 tasks——
Arc Institute Virtual Cell ChallengeVirtual Cell2025Perturbation PredictionActive—Statistical baselines competitive with AI——
scPerturbVirtual Cell2024Data Resource/QCActive—E-distance metric—scPerturBench
scPerturBenchVirtual Cell2025Perturbation PredictionActive—No method consistently wins; linear baselines competitive——
GEARSVirtual Cell2023Perturbation PredictionActive—40% higher precision than baselines (claimed); linear models outperform (2025)——
PerturBenchVirtual Cell2024Perturbation PredictionActive—Simple models outperform complex——
BEELINE (GRN inference)Virtual Cell2020GRN InferenceActive—Close to random predictor in many cases——
dynverse (Trajectory inference)Virtual Cell2019Trajectory InferenceActive—Varies by topology——
PubMedQABio NLP2019MCQSaturated72 mo81.6%78%MedHELM, BigBIO
MedQA (USMLE)Clinical2021MCQSaturated42 mo96.5%87%MedXpertQA, HealthBench
MedMCQAClinical2022MCQActive—~75-80%——
MMLU-BioBio NLP2020MCQSaturated48 mo93%+89.8%MMLU-Pro
MMLU-Pro MedicalBio NLP2024MCQSaturated18 mo90.1%——
BLURBBio NLP2020Multi-task NLPNearing72 mo82.91 BLURB score—BigBIO
BigBIOBio NLP2022FrameworkActive—N/A (126+ datasets)——
BioASQBio NLP2013Semantic QAActive—F1 ~0.58; yes/no >80%——
BC5CDR-Chemical NERBio NLP2015NERSaturated84 moF1 94.2%—BioRED
BC5CDR-Disease NERBio NLP2015NERNearing132 moF1 ~90%——
NCBI-Disease NERBio NLP2014NERSaturated96 moF1 ~91%IAA ~87%—
JNLPBA NERBio NLP2004NERSaturated204 moF1 ~79.6%——
BC2GM NERBio NLP2007NERNearing228 moF1 ~90.9%——
ChemProt REBio NLP2017Relation ExtractionNearing108 moF1 90.8%——
DDI REBio NLP2013Relation ExtractionSaturated120 moF1 83.3%—MultiADE
GAD REBio NLP2004Relation ExtractionSaturated204 moF1 ~85%——
HoC (Hallmarks of Cancer)Bio NLP2016ClassificationNearing120 moF1 ~90.3%——
BioREDBio NLP2022Relation ExtractionActive—Varies——
MedNLIBio NLP2018NLINearing96 mo~82% accuracy—BioNLI
LINNAEUS NERBio NLP2010NERSaturated132 moMid-to-high 90s%——
GPQA Diamond (Bio)Science QA2023MCQSaturated36 mo94.1%67%HLE, ATLAS
HLE (Bio/Medicine)Science QA2025Short Answer/MCQActive—44.7% overall; ~22% bio/med98%—
ATLAS (Bio)Science QA2025Open-endedActive—Varies——
MedHELMClinical2025Multi-task ClinicalActive—66% win-rate (DeepSeek R1)——
MedAgentBenchClinical2025Agent WorkflowActive—69.67%——
HealthBenchClinical2025ConversationalActive—60% (o3); Hard: 32%——
MedCalc-BenchClinical2024CalculationActive—50.9%——
MIMIC-III BenchmarksClinical2017PredictionNearing108 moAUROC ~0.94 (mortality)—MIMIC-IV, YAIB
MIMIC-IV BenchmarksClinical2020PredictionActive—Hospitalization AUROC ~0.87——
EHRSHOTClinical2023Few-shot PredictionActive—Varies across 15 tasks——
FHIR-AgentBenchClinical2025Agent WorkflowActive—Varies——
LiveMedBenchClinical2025Dynamic MCQActive—Varies (refreshes weekly)——
MedXpertQAClinical2025MCQActive—Varies (4,460 Qs, 17 specialties)——
DiagnosisArenaClinical2025Diagnostic ReasoningActive—45.82%——
MedS-BenchClinical2024Multi-taskActive—Varies across 39 datasets——
Script Concordance TestingClinical2025Probabilistic ReasoningActive—Varies——
Doctorina MedBenchClinical2026Agent SimulationActive—Varies (>1,000 cases)——
AgentClinicClinical2025Diagnostic AgentActive—Varies (120 NEJM cases)——
ACI-BENCHClinical2023Note GenerationActive—MEDCON 57.78——
PhysioNet 2019 (Sepsis)Clinical2019PredictionActive—Varies (104 teams)——
JAMA Clinical ChallengeClinical2024MCQActive—88.6% (o1-preview on 70 cases)——
MedBench v4 (China)Clinical2025Multi-taskActive—Agent 85.3/100 (Claude Sonnet 4.5)——
AfriMed-QAClinical2024MCQActive—Varies (15K+ Qs, 32 specialties)——
MMedBenchClinical2024Multilingual MCQActive—Varies across 6 languages——
TAC ADRBio NLP2017ExtractionNearing108 moF1 ~85.2%—MultiADE
MultiADEBio NLP2024ExtractionActive—Varies across 6 ADE datasets——
SMM4HBio NLP2017Social Media MiningActive—ADR detection F1 ~0.65-0.70——
WMDP-BioBiosecurity2024MCQSaturated7 mo87%60.5%VCT, ABC-Bench
VCTBiosecurity2025Multimodal QAActive—43.8% (o3)22.1% (expert virologists)—
ABC-BenchBiosecurity2025Agent WorkflowActive—53% (Grok 3)24% (PhD experts)—
ABLEBiosecurity2025Agent WorkflowActive—7/8 low-level tasks——
BBG Framework / B3Biosecurity2025Open-endedUnknown—Pilot phase——
BioLP-BenchBiosecurity2024Protocol Error DetectionActive—34% (o4-mini)~38% (expert avg)—
Scale AI FORTRESS (CBRNE)Biosecurity2025Adversarial SafetyActive—Varies——
Scale AI PropensityBenchBiosecurity2025Agent SafetyActive—Varies (5,874 tasks)——
TroubleshootingBenchBiosecurity2025Protocol QAActive—No model exceeds 80th-pct expert (36.4%)80th-pct expert: 36.4%—
LAB-BenchAgentic Bio2024Research WorkflowSaturated18 mo89%; several subtasks at ceiling~79% (expert on ProtocolQA)LABBench2
LABBench2Agentic Bio2026Research WorkflowActive—26-46% drop vs LAB-Bench——
BixBenchAgentic Bio2025BioinformaticsActive—17% (Claude 3.5 Sonnet); 9% (GPT-4o)——
ScienceAgentBenchAgentic Bio2024Code GenerationActive—42.2% (o1-preview + self-debug)——
BioCoderAgentic Bio2024Code GenerationNearing24 mo~50% Pass@1 (GPT-4)—BixBench
MedAgentGymAgentic Bio2025Training EnvironmentActive—N/A——
LitQA2 / PaperQA2Agentic Bio2024Literature QANearing24 mo~90% (PaperQA2)~67% (human experts)—
BioProBenchAgentic Bio2025Protocol UnderstandingActive—Varies——
CheXpertMedical Imaging2019ClassificationNearing84 moAUC ~0.942.6/3 radiologistsCheXpert Plus
MIMIC-CXRMedical Imaging2019Report GenerationActive—RadCliQ-v1 0.92——
NIH ChestX-ray14Medical Imaging2017ClassificationNearing108 moAUC ~0.85-0.88—CheXpert, MIMIC-CXR
PathVQAMedical Imaging2020Visual QAActive—50-65%—GEMeX
VQA-RADMedical Imaging2018Visual QAUnknown—79.2%——
SLAKEMedical Imaging2021Visual QAActive—~78.7%——
OmniMedVQAMedical Imaging2024Visual QAActive—LVLMs struggle——
ISIC ChallengesMedical Imaging2016Classification/SegmentationNearing120 moExceeds cliniciansClinician level3D TBP (2024)
HAM10000Medical Imaging2018ClassificationNearing96 mo~96%+ accuracy—DermaMNIST-E
Camelyon16Medical Imaging2016ClassificationSaturated72 moAUC 0.994Pathologist AUCCamelyon17, Camelyon+
Camelyon17Medical Imaging2017ClassificationActive—Kappa ~0.89—Camelyon+
PANDA (Prostate Gleason)Medical Imaging2020ClassificationActive—QWK ~0.93+——
PCamMedical Imaging2018ClassificationNearing96 mo~97%+ accuracy——
BraTSMedical Imaging2012SegmentationActive—Dice ~0.88 whole tumor—BraTS 2025 Lighthouse
Medical Segmentation DecathlonMedical Imaging2018SegmentationActive—nnU-Net variants lead——
TotalSegmentatorMedical Imaging2023SegmentationNearing36 moAvg Dice >0.90 for major organs——
AMOSMedical Imaging2022SegmentationActive—Varies (15 organs)—AMOS-MM (2024)
CT-BenchMedical Imaging2025Multimodal QAActive—61.8%——
APTOS Diabetic RetinopathyMedical Imaging2019ClassificationNearing84 moQWK 0.967——
EyePACS DRMedical Imaging2015ClassificationSaturated48 moAUC ~0.99——
RSNA Pneumonia DetectionMedical Imaging2018Object DetectionActive—Varies (1,400+ Kaggle teams)——
ReXrankMedical Imaging2024Report GenerationActive—1/RadCliQ-v1 0.98 (ReXGradient)——
RadGraphMedical Imaging2021Entity/Relation ExtractionActive—Micro F1 0.82——
ReXVQAMedical Imaging2025Visual QAActive—83.24% (MedGemma)——
MatBench DiscoveryMolecular/Drug Discovery2025Crystal StabilityNearing12 moF1=0.924——
OC20Molecular/Drug Discovery2020Catalyst PredictionActive—EquiformerV2 leads—OC22, OC25
MedPerfClinical2023Federated BenchmarkingActive—N/A (platform)——
FairMedQAClinical2025Fairness EvaluationActive—3-19 pct point accuracy disparity——
MEDECClinical2025Error Detection/CorrectionActive—Varies——
MedThink-BenchClinical2025Reasoning EvaluationActive—Varies (500 QA pairs)——
CLINICClinical2025TrustworthinessActive—DeepSeek-R1-LLaMA best robustness——
MedAgentsBenchClinical2025Multi-agent ReasoningActive—92.6% (TeamMedAgents)——